#2192 Crash sometimes serving IQ3_S

Descrição

Crash during output when serving IQ3_S

dumps:
[strata-stall-21128.dmp](https://github.com/user-attachments/files/33359520/strata-stall-21128.dmp)
[strata-stall-28072.dmp](https://github.com/user-attachments/files/33359521/strata-stall-28072.dmp)

logs:
```
strata-vision: model files loaded in 4.4 s
strata-vision: warmed up at 1024 image tokens in 0.9 s
[strata] 2026-10-13 06:17:10 engine started: C:\Strata-0.1.43\engine\strata.exe --serve --pack C:\Strata-data\packs\iq3_s --native C:\Strata-data\models\IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf --ple-gguf C:\Strata-data\models\IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf --expert-profile C:\Strata-0.1.43\data\expert-profile.bin --expert-cache auto --prefill auto --spec 4 --mtp C:\Strata-data\mtp\rt --max-context 262144 --kv int8 --kv-resident 32768 --vision --vram-reserve-mib 700 --control-vector-scaled C:\Strata-0.1.43\data\experimental-speed-projection\Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer --pcie-frac 0.55 --spec-min-p 0.70 --pool-workers 17 --batch 4 --layer-split auto
strata generate: memory guard: limit 511.9 GiB (source: meminfo), current 408.9 GiB, reclaimable cache credited 0.0 GiB; available 103.0 GiB (MemAvailable 103.0 GiB), floor 2560 MiB
strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)
strata generate: this CPU has no AVX-512: the expert kernels run on AVX-2 (multi-token for the i-quant gate/up rows)
strata generate: native pack: C:\Strata-data\packs\iq3_s experts (largest blob 2.66 MB), token embedding IQ4_XS in mapped host memory (322 MiB, 2.0 s)
strata generate: 1466 MiB of weights loaded from C:\Strata-data\packs\iq3_s in 1.1 s (302 canonical tensors skipped: served natively)
strata generate: 300 native projection matrices, 2018.88 MiB of weights, in 11.1 s
strata generate: control vector mode = project, dir = per-layer, layers 4..44 (41 steered)
strata generate: PLE direct read probe: 25156 rows/s (fine)
strata generate: the SSD is kept awake while rows are read: one page of the table after 100 ms without a read, until 60 s after the last request (STRATA_SSD_KEEPALIVE=0 turns it off)
strata generate: layer split: CUDA1 holds its weights; 17.15 GiB free (its session follows the split search)
strata generate: profile C:\Strata-0.1.43\data\expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: layer split auto: CUDA0 68 SMs at 1.54 GHz -> 0.69 ms per layer, 16.47 GiB free before its session carve
strata generate: layer split auto: CUDA1 68 SMs at 1.54 GHz -> 0.69 ms per layer, 15.40 GiB free before its session carve
strata generate: layer split auto: CUDA0 link 5.9 GB/s -> miss cost x3.37
strata generate: layer split auto: CUDA1 link 11.4 GB/s -> miss cost x1.75
strata generate: layer split auto: K=18 - predicted 37.6 ms per decode window; the caches hold 16487 of 24576 profiled pairs (~98.7% of the routed mass), best of 46 placements
strata generate: layer split auto: NEAR TIE - K=21 is predicted 37.8 ms per decode window, within 0.7% of K=18 (37.6 ms); the cards differ, so the prediction may not rank them (it prices decode, not the prompt). To try the other: --layer-split 21
strata generate: KV streaming: 32768 of 262144 cells per QSA layer in VRAM, the K/V in 1.03 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of C:\Strata-data\models\IQ3_S\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf
strata generate: layer split: CUDA1 holds its weights, session [18, 48) and the head; 16.02 GiB free
strata generate: --batch-groups auto: 2 groups of 2 slots
strata generate: --batch: 4 slot sessions on CUDA0 (0.37 GiB each); 15.30 GiB free
strata generate: --batch: 4 slot sessions on CUDA1 (0.65 GiB each); 13.42 GiB free
strata generate: --batch 4: the slot sessions take 1.48 GiB of VRAM on CUDA0 that the expert cache would otherwise hold
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.94 s (835 MiB/s)
strata generate: GPU 0: NVIDIA GeForce RTX 2080 Ti, compute capability 7.5
strata generate: multi-GPU under WDDM: at most 8 GiB of the expert arena is pinned (STRATA_ARENA_PIN_GIB changes it)
strata generate: expert arena read unbuffered (0 of 16 probe reads from the file cache; 39.3 GiB available, 51.1 GiB of files)
strata generate: expert arena: large pages are resident without a lock; cudaHostRegister limited to 8 GiB by the engine (multi-GPU under WDDM, or remote experts); 8 slices pinned (7 GiB); large pages (2097152 B)
strata generate: loaded 46.84 GiB at 0.43 GiB/s
strata generate: CPU pool tasks/phase: 54 (automatic), participating threads: 18
strata generate: 17 pool workers on logical processors 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17, the host thread on 0 (draining too) (--host-core first)
strata aux cpus: not available on this system
strata generate: expert cache auto: 15.30 GiB free, 700 MiB reserved (+0 MiB for the draft head) -> 5894 slots
strata generate: only 0 MiB free once the slots are written (reserve 700 MiB); shrinking the expert cache
strata generate: only 0 MiB free once the slots are written (reserve 700 MiB); shrinking the expert cache
strata generate: expert cache 7059 slots, 12.61 GiB of VRAM; policy is
strata generate: the GPU computes the experts in the cache; it rounds differently from the CPU,
                 so a reply can differ slightly from a run without the cache (same quality:
                 bench/results/2026-09-27-cache-parity).
                 PROFILE, ranked by routing frequency, no eviction.
strata generate: pre-filled 7059 of 7059 slots from the profile in 3.2 s (4272 MB/s); slot 0 verified
strata generate: layer split, CUDA1: 12.60 GiB free of 22.00, room for experts 11.82 GiB
strata generate: layer split: CUDA1 runs layers 18-47, expert cache 5993 slots (11.82 GiB), 5993 of its 15360 profiled pairs; slot 0 verified
strata generate: layer split: CUDA0 runs layers 0-17
strata generate: layer split auto: L2 probe: the split is startable on the live free figures
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 17 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.43)
strata generate: token graph hit path: 13052 resident experts, decided on the device
strata serve: Windows commit: 8.6 GiB left of a 527.9 GiB limit (RAM + page file)
strata serve: prompt chunk auto: 8192 tokens, a 96-slot ring
strata serve: the prompt path borrows 2034 CUDA0 cache slots (3.66 GiB)
strata serve:   CUDA1 prompt path borrows 1865 of its 5993 slots (3.66 GiB)
strata hc: CUDA1: the hyper-connection read runs as staged (the norm per token and stream, the down projection's activations staged ahead by cp.async); checked bit for bit against the plain read on this card (STRATA_HC_SPLIT=1 or 0 for the earlier ones)
strata verify: window up to 6 tokens, 75.8 MiB of device buffers
strata serve: layer split: layers 0-17 (CUDA0), 18-47 (CUDA1), one hand-off per window
strata hc: CUDA0: the hyper-connection read runs as staged (the norm per token and stream, the down projection's activations staged ahead by cp.async); checked bit for bit against the plain read on this card (STRATA_HC_SPLIT=1 or 0 for the earlier ones)
strata verify: window up to 6 tokens, 75.8 MiB of device buffers
strata mtp: draft head over 106299 tokens (212.9 MiB)
strata verify: batch windows of up to 4 sequences (layers [0, 18))
strata verify: batch windows of up to 4 sequences (layers [18, 48))
strata serve: 740 MiB of VRAM free with everything loaded
strata verify: capturing the 4-token window (671 MiB of VRAM free)
strata verify: captured the 4-token window (upload no error, sync no error)
strata verify: capturing the 4-token window (153 MiB of VRAM free)
strata verify: captured the 4-token window (upload no error, sync no error)
strata verify: capturing the 1-token window (607 MiB of VRAM free)
strata verify: captured the 1-token window (upload no error, sync no error)
strata verify: capturing the 1-token window (82 MiB of VRAM free)
strata verify: captured the 1-token window (upload no error, sync no error)
strata serve: prompt 270 tokens = 0 reused + 270 read in 3386 ms (79.7 tok/s), 1 generated in 67 ms (14.9 tok/s), drafts accepted 0 of 0, 1 checkpoints
strata serve: decode expert cache hit rate: 82.7% (396 hits / 479 lookups); 1 more read by the GPU over PCIe or from another GPU (0.2% of all 480 routed)
strata serve: KV streaming: 79.82% of 1348 block reads hit VRAM, 1.1 MiB read from RAM
strata batch: slot 0 takes 271 tokens (copied in 163.3 ms)
strata serve: prompt 271 tokens = 270 reused + 1 read in 0 ms (3369.3 tok/s), 1 generated in 25 ms (40.7 tok/s), drafts accepted 0 of 0, 1 checkpoints
strata serve: decode expert cache hit rate: 88.8% (426 hits / 480 lookups)
strata serve: KV streaming: 83.21% of 1620 block reads hit VRAM, 1.1 MiB read from RAM
strata verify: captured the batch window over slots 0
strata verify: captured the batch window over slots 0
strata verify: capturing the 6-token window (633 MiB of VRAM free)
strata verify: captured the 6-token window (upload no error, sync no error)
strata verify: capturing the 6-token window (52 MiB of VRAM free)
strata verify: captured the 6-token window (upload no error, sync no error)
strata batch: slot 2 takes 6474 tokens (copied in 266.3 ms)
strata serve: prompt 6474 tokens = 0 reused + 6474 read in 11100 ms (583.2 tok/s), 1 generated in 26 ms (38.3 tok/s), drafts accepted 0 of 0, 2 checkpoints
strata serve: decode expert cache hit rate: 84.8% (407 hits / 480 lookups)
strata serve: KV streaming: 89.28% of 34920 block reads hit VRAM, 15.1 MiB read from RAM
strata verify: captured the batch window over slots 2
strata verify: captured the batch window over slots 2
strata batch (pipelined, 2 groups of 2): 323 group-steps, 323 rows in 20871 ms = 15.5 rows/s (admissions included)
strata batch: slot 2 gave back 6636 tokens of this conversation (all it holds) in 219.6 ms
strata verify: capturing the 2-token window (574 MiB of VRAM free)
strata verify: captured the 2-token window (upload no error, sync no error)
strata verify: capturing the 2-token window (137 MiB of VRAM free)
strata verify: captured the 2-token window (upload no error, sync no error)
strata verify: capturing the 3-token window (634 MiB of VRAM free)
strata verify: captured the 3-token window (upload no error, sync no error)
strata verify: capturing the 3-token window (77 MiB of VRAM free)
strata verify: captured the 3-token window (upload no error, sync no error)
strata serve: prompt 6637 tokens = 6636 reused + 1 read in 220 ms (4.5 tok/s), 41 generated in 1256 ms (32.7 tok/s), drafts accepted 20 of 26, 0 checkpoints
strata serve: decode expert cache hit rate: 86.6% (19496 hits / 22502 lookups); 58 more read by the GPU over PCIe or from another GPU (0.3% of all 22560 routed)
strata serve: KV streaming: 95.14% of 96548 block reads hit VRAM, 18.9 MiB read from RAM
strata batch: slot 0 takes 6678 tokens (copied in 230.0 ms)
strata serve: prompt 6678 tokens = 6677 reused + 1 read in 1 ms (940.6 tok/s), 1 generated in 27 ms (37.5 tok/s), drafts accepted 0 of 0, 0 checkpoints
strata serve: decode expert cache hit rate: 89.1% (426 hits / 478 lookups); 2 more read by the GPU over PCIe or from another GPU (0.4% of all 480 routed)
strata serve: KV streaming: 95.22% of 98604 block reads hit VRAM, 19.0 MiB read from RAM
strata batch: slot 2 takes 266 tokens (copied in 166.4 ms)
strata serve: prompt 266 tokens = 0 reused + 266 read in 2560 ms (103.9 tok/s), 1 generated in 24 ms (42.4 tok/s), drafts accepted 0 of 0, 1 checkpoints
strata serve: decode expert cache hit rate: 83.3% (399 hits / 479 lookups); 1 more read by the GPU over PCIe or from another GPU (0.2% of all 480 routed)
strata serve: KV streaming: 79.82% of 1328 block reads hit VRAM, 1.1 MiB read from RAM
strata batch: slot 1 takes 6468 tokens (copied in 2